Sampling-Based Aircraft Path Planning with Soft Actor-Critic

This paper investigates the usage of reinforcement learning for a global path planning task in terminal airspace, specifically for training a policy that can generate paths from any given position in the airspace to the runway. To do this, the Soft Actor-Critic (SAC) algorithm is trained on a simplified version of the Dutch airspace and compared to the solutions generated by the Dijkstra algorithm for varying discretization resolutions. SAC, which uses a Gaussian distribution for the action policy, has previously been shown to be successful in other global planning tasks in continuous environments. However, evaluating the policy by following the mean of the learned distribution, which is the standard evaluation method, may yield suboptimal performance when dealing with complex cost functions that deviate from a normal distribution. To address this, the paper proposes and evaluates a sampling-based strategy, which generates an ensemble of paths by sampling from the learned policy distribution. These three methods: mean-based SAC, Dijkstra and sampling-based SAC, are then tested on a bi-criterion cost function which includes both fuel and noise emissions in varying ratios. It was found that Dijkstra outperforms mean-based SAC for all cost ratios at the best discretization resolution, regardless of the neural network architectures used. However, sampling-based SAC results in consistently lower costs than both Dijkstra and mean-based SAC, particularly for the more complex cost functions that have a higher focus on noise mitigation. These findings highlight some limitations in mean-based evaluation for distribution models and indicate potential performance benefits that can be obtained with better-tailored evaluation strategies.